Relay framework beats 11/12 benchmarks in cost-efficient LLM evolution

LLM-driven evolution—using language models to search programs and discover algorithms—often relies on strong models for long runs, which is expensive. A cheaper strategy is to mix weak and strong models under a fixed inference budget, but existing approaches allocate models per query, ignoring that evolution is stateful: each candidate mutates the population for the next step. In a new arXiv paper, Sichun Luo and colleagues analyze evolutionary trajectories and find that progress is heavily front-loaded: cheap models recover much of the early gains of strong models at lower cost, though early performance is noisy.

To exploit this, they propose Relay, a training-free method that treats budget allocation at the population level. A cheap model explores multiple short trajectories, with a bandit scheduler deciding when to hand off based on 'Relay Gain'—the marginal improvement of a compact, quality-diverse candidate bank. These curated candidates then initialize a shared strong-model population for refinement. Across four benchmarks and three budgets, Relay achieved the highest mean score in 11 of 12 settings, beating competitive baselines. The work suggests that in stateful search, you should route resources around the population's needs, not individual calls, offering a practical blueprint for cutting inference costs in evolutionary AI systems.